Currently, through the Discourse Automation plugin, you can generate responses to topics based on keywords that get detected. This option is available as part of Auto Responder script and can be configured through the List of word/answer pairs option. Additionally, you can use AI triage to provide canned responses on topics based on LLM prompts, its results and post context.
The feature request idea here is to put an AI spin using an LLM to decipher what’s written through customizable prompts instead of relying on keywords. The LLM can then intelligently respond instead of just using canned responses.
Curious to hear if other people are interested in something like this and what you would use it for. Example use-cases are always appreciated!
Absolutely! There are countless use cases for this. Think of some basic support bot: user posts in “Help” category because they don’t know how to use some Discourse functionality. Then there can be a designated persona which automatically responds to each new topic in that category. Coupling this with RAG, tools, and detailed prompt engineering, you can get quite creative.
Hypothetically each category could have its own “category expert bot” then.
The idea is that you can select a tool to do the triage, it in turn can create topics, like stuff, do whatver you want. I hope to continue work on it soon.
This came in internally as a request, so exploring the idea based on community examples and use-cases, which Sam is now working on
An example for me would be nudging users to post in the right place. I.e, if a topic seems like a mismatch with the category, instead of making that change automatically, perhaps replying in the topic suggesting other categories
Our use case for this is that when a user posts a question in our Discourse we would like AI to leverage our trained model to give a first response to the first post. We’re trying to give the user a quick, educated response with AI, with us following up if needed.
I do love the idea of having an AI category expert per category.
Will work on documenting tomorrow. We have what I would call mode 1 and 2 implemented:
Ultra trivial, select target and unconditionally reply (category / tag / pm group etc)
Triage + Responder, which allows you to triage and then respond dynamically with llm triage
I also implemented (3) mega sophisticated mode - which is still in progress and disabled for now.
Triage using AI tool
This is unfinished, but when done will allow for some very fancy stuff like triaging using an LLM and then routing to the correct persona and doing other extremely complicated things.
Personally, I don’t like the idea of auto-responding due to the risk of annoying our users too much (but maybe we’re spoiled with a decent reply rate to our new topics from our users).
However, I like this use case the most:
We actually have something like this automated with an API for years now, where we reply with a canned reply after 24 hours with no replies on a topic. Being able to have this canned reply augmented with LLMs would be fantastic (as long as there is a toggle to conditionally reply only on unanswered topics after x hours).
Whispers (which is a mode we support) is one way around this, use it to assist moderators.
I also think that some fancy flows can reduce a lot of risk here (it will have to wait for tools)
Craft a response
Ask LLM to gauge how much value the response adds from 1 to 10
If 8 or higher then post response
I love this idea, it will require a new type of automation trigger, going to have to think about it. This type of “trigger on topics in category X that have not gotten any response after a day - check every 10 minutes” is not a trigger we have at the moment… we need to figure out how to add it to automation.
We have a “stalled topic” trigger but I don’t think it has the ability to “bypass” on “has a reply”, we may need to expand it.
Additionally for API integrations a “just answer this topic with this persona” action would be pretty amazing and very very easy to build.
